AISD software
AISD software uses a cascaded Residual Network (ResU-net) algorithm to segment aneurysms on the original axial images in head and neck CT angiography, including ResU-net1 and ResU-net2 modules. The ResU-net1 network is used for aneurysm detection, with its encoder using 5-layer down-sampling and the decoder using 5-layer up-sampling. Input channel number is 1, input original axis image, output channel number is 2. If the size of the output result is the same as the input size, the number of channels becomes 2. Each voxel point in channel 1 represents the probability of an aneurysm. When the probability value is greater than 0.5, it is considered that the current voxel point is an aneurysm, which serves as the standard for AI diagnosis of aneurysms. Each voxel point in channel 2 represents the length, width, and height information of the aneurysm when the current voxel point is the center point of the aneurysm. It will cut the detected aneurysm into 48 × 48 × 48 pixel data cubes and input them into the ResU-net2 network for aneurysm segmentation. The ResU-net2 network encoder uses 3 layers of down-sampling, while the decoder uses 3 layers of up-sampling. After outputting the segmentation result of the aneurysm, the generated aneurysm segmentation result is restored to the corresponding position in the image through post-processing. Using Digital Subtraction Angiography as the gold standard, the sensitivity, specificity, and accuracy of diagnosing intracranial aneurysms were 95.9%, 92.4%, and 95.4%, respectively. The automatic detection of fractures is based on V-Net. V-Net provides a 3D convolutional neural network architecture for extracting fracture features and locating images. The neural network first extracts corresponding features through the image feature compression network path, and then recovers to a three-dimensional matrix of the same size as the input through a decompressing network symmetric to the compression network. The software processing result interface includes horizontal axis images with rib fracture markers, curved planar reconstruction images, and VDR images to facilitate diagnostic physicians in observing the fracture situation detected by AI. The software processing result page includes horizontal axis images with rib positioning labels and fracture site markers. Users can simultaneously perform post-processing of Multi Planar Reconstruction (MPR) and VDR in the software to observe rib fractures detected by AI from multiple perspectives. When diagnosing rib fractures, the software can also automatically represent the type of fracture, such as displaced fractures, non-displaced fractures, and old fractures. Post-processing technology of MPR and 3D is used to help identify fractures and determine the nature of fractures. Its detection rates for dislocation fractures, non-dislocation fractures, and old fractures are 96.04%, 97.02%, and 98.79%, respectively. The total missed diagnosis rate for different types of rib fractures is 2.53%. Pulmonary nodule recognition utilizes U-net networks for lung segmentation. The networks separate the lung area from the background and other organs, retaining only CT images containing lung information. This program utilizes Retina-Net network for lung tissue nodule detection, locates all possible candidate regions of lung nodules in lung CT images, and outputs their positions and confidence levels. It then uses the Non-Maximum Suppression algorithm to screen candidate regions for pulmonary nodules, removing duplicate or low confidence candidate regions and preserving the final pulmonary nodule detection results. After cropping and scaling each candidate nodule, the program obtains a uniformly sized small image block as input to the segmentation network. It sets the size of small image blocks to 64 × 64 to ensure consistent input size for the segmentation network, while avoiding the impact of excessively large or small image blocks on the segmentation results. DeepLabv3 + network is used for lung nodule segmentation, which separates the lung nodules in small images from the background and outputs the segmentation results. At the same time, conditional random field algorithm is used to optimize and refine the lung nodule segmentation results. Dense-Net network is used for lung nodule classification, which distinguishes lung nodules in small image blocks from the background and outputs classification results. The cross-entropy damage function and accuracy are used as optimization objectives and evaluation indicators for the classification model, respectively, to measure the prediction error and accuracy of the model for the category of pulmonary nodules. The detection method for pulmonary nodules has sensitivity, specificity, and accuracy indicators of 95%, 98%, and 96%, respectively.
Design
The AISD software based on VDR (Shukun Internet Sci & Tec Co., Ltd., Beijing) was installed in January 2020 and used in the practical teaching of the 2017 grade students, while the previous students used traditional teaching (video: AISD operation video and function demonstration). The data in the form of Digital Imaging and Communications in Medicine (DICOM) from the Picture Archiving and Communication System (PACS) of our hospital was imported to the software. The 3D images were displayed by endowing the tissues with colors via the software. Meanwhile, trans axial, coronal and sagittal images were displayed. The students can magnify, minify, revolve, adjust windows, and hide/display images. The lesions were marked, and the students’ diagnostic suggestions were given. Then the images were analyzed with AI and the lesions were re-marked, which were pushed to the students (Fig. 1). The two groups were both taught by the same imaging teacher who had 10 years of clinical experience. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. This study was approved by the Ethics Committee of the Second Affiliated Hospital of Qiqihar Medical College (No. 20,190,017) and obtained the written informed consent of all participants.
Research process and teaching implementation flowchart
Participants
In January 2020, AISD software based on VDR technique (Shukun Network Technology Co., Ltd., Beijing) was installed and applied in practical teaching of image reading. The student group was based on this time point as the boundary. Prior to this date, all 2016 students who came to the hospital for internships did not use AISD software and were designated as the control group. After this date, all students who came to the hospital for internships in 2017 would receive AISD software learning and be designated as the experiment group. This comparative study using previous and subsequent students avoided psychological imbalance between two groups of students and bias in the final data results. Totally 41 students majoring in clinical medicine in 2017 from Qiqihar Medical University were enrolled as the experiment group, including 18 males and 23 females and aged 19–25 years old (21.73 ± 1.41 years). The comprehensive measuring and evaluating scores at 1 year before the enrollment were 84.73 ± 11.72. Then 43 students majoring in clinical medicine in 2016 were enrolled as the control group, including 19 males, 24 females and aged 20–27 years old (22.35 ± 1.65 years). The comprehensive measuring and evaluating scores at 1 year before the enrollment were 83.98 ± 9.58. No significant difference was found in the baseline data between the two groups (P > 0.05). The practical course of teaching was Medical Imaging with 20 classes. The contents, hours and textbook of teaching were all the same between the two groups. The teachers were totally the same between the two groups.
Experimental procedure
Implementation of teaching in the experiment group: The medical history, physical examination, and imaging of typical clinical cases were collected as per the syllabus. The image data of computerized tomography (CT), magnetic resonance (MR) and enhanced scan from the PACS were imported to the software. Then transaxial, coronal and sagittal images were generated, which clearly showed the relationship between lesions and the peripheral anatomic structures. The positions, sizes and properties of the lesions were analyzed via AI (Fig. 2). Before the teaching, the students were informed about the concrete cases. During the explanation, the students were guided stepwise to analyze the cases. Then questions were raised for the students to discuss, and finally the teacher summarized the lessons. During the teaching, the images of the cases were sent to coronal and sagittal reconstruction and VDR on the software, so the students can revolve and observe the sizes, extents, and relationships with adjacent blood vessels from multiple perspectives. After the students marked the lesions, they can submit it to the software. Then after image analysis via AI, text annotations were generated (Fig. 3), which were pushed again to the students. Each student can search the cases of interest in the software for learning. Each case was annotated with diagnostic imaging characteristics via AI, which helped with learning. Taking respiratory system practice image reading as an example, AISD software automatically identifies suspicious lesions in the entire lung of selected cases, and can label and report the location, size, morphology, nodule type, CT value, lung imaging reporting and data system score, likelihood of nodule malignancy, etc. Through AI diagnosis results, detailed information of corresponding lesions can be viewed. At the same time, by using the mouse and keyboard to zoom in, out, and rotate the VDR stereoscopic image, the surface morphology of the lesion nodule and its relationship with surrounding pulmonary blood vessels and bronchi can be observed more clearly (Fig. 4), and the subtle changes in its own structure and surrounding areas can be more fully displayed.
Software images of respiratory cases the software can display transaxial and sagittal images. The AI circles marked the position (dorsal inferior lobe of right lung), size (1.4 cm×1.2 cm), volume (1.782 cm3) and density (-104 Hu) of the lesions, and presented the suggestions on the follow-up to the lesions
Software images of head and neck vascular cases. The software provides VDR stereoscopic images (position of aneurysm marked by AI) that can be revolved and magnified, and offers horizontal images (positions of lesions marked on AI circles), coronal images (AI measured aneurysm size = 4.8 mm×4.3 mm×2.6 mm, volume = 65.9 mm3), and stretched camber images of blood vessels at the lesions. It also displays a complete relationship between blood vessels and lesions on the plane view
VDR screenshot of respiratory system case. In three-dimensional images, lesions were displayed in the form of pink and yellow nodules, and the distribution and surface morphology of lesions in the lobes and segments of the lungs, as well as the relationship between surrounding pulmonary blood vessels and bronchi, were clearly displayed
Implementation of teaching to the control group: The traditional teaching method was used in the control group. Before the teaching, the students were informed about the concrete cases, and the cases were explained to them via PPT. During the explanation combine the students were directed stepwise to analyze the cases. Images were played via PPT, and real image films were presented to the students. Then questions were raised for the students to discuss, and finally the teacher summarized the lessons. According to the syllabus, the teacher chose typical images from the PACS, and made PPT of the cases. During the teaching process, students analyzed and learned abnormal imaging signs through continuous observation at the two-dimensional image level in PACS, and difficult images were assisted by teachers to answer.
Evaluation of teaching effect
Final exam
Theoretical knowledge exam and image reading skill exam were adopted. The total score of theoretical exams was 100, including 20 for term explanation, 30 for brief questions, 10 for discussion, and 40 for single choices. In the image reading skill exam, 50 reading images for single choices of diagnosis were made as per the syllabus, which included X-ray diagnosis [20], CT diagnosis (60), and MR diagnosis [20]. The time of the exam was 60 min. All images were real cases chosen from PACS.
Academic self-efficacy scale
This scale was compiled by Liang Yusong from Central China Normal University [13]. The academic self-efficacy was divided into two independent dimensions, including learning ability self-efficacy, and learning behavior self-efficacy [7]. The learning ability self-efficacy (11 terms) refers to the evaluation whether a student can accomplish the academic affair, achieve satisfactory result, and avoid failure. The learning behavior self-efficacy (11 terms) refers to the evaluation whether a student thinks his/her learning behavior can achieve the learning objective. Each term has 5 grades (1–5 scores): fully unqualified (1 score), very unqualified (2 scores), a bit qualified (3 scores), very qualified (4 scores) and fully qualified (5 scores).
Self-directed learning scale
This scale involves 5 dimensions: consciousness of learning, learning behavior, learning strategy, learning evaluation, and interpersonal skills, and each dimension has 12 terms. Each term has 5 grades (1–5 scores): never (1 score), seldom (2 scores), occasionally (3 scores), often (4 scores) and always (5 scores). Each student chose the most appropriate answer according to his/her thoughts and feelings about learning. A larger score means the student has stronger self-directed learning ability.
Statistical analysis
Data were analyzed on SPSS 18.0. Shapiro Wilk test for normal distribution test was adopted. The Quantitative data were expressed as Mean ± Standard Deviation (SD), and Shapiro-Wilk test was used for normal distribution test. When the data were normally distributed and showed equal variance within groups, student t test was performed; otherwise, Mann-Whitney U test was used. Qualitative data are expressed as percentages and Chi-square tests were used for comparison. P < 0.05 was considered to be statistically significant.
